VC-Attention Speeds Up Low-Bit Attention for Video Generators on GPU

The authors of VC-Attention: Value Smoothing and Softmax Casting for Low-bit Attention propose a training-free low-bit attention method for Diffusion Transformers. It solves two problems: quantization errors caused by outliers in value tokens and the slow softmax between matrix multiplications. The implementation is stated for B200, B300, H200, RTX PRO 6000 and RTX 5090; the authors have not released the code.

What it means

The work directly concerns video generators from our AI video section. Among the tested models is HunyuanVideo from Tencent. This is an open source text-to-video generator that, as described by the author, requires 45—60 GB of VRAM and runs on CUDA. The code and model weights are under the Tencent Hunyuan Community License: commercial use is permitted worldwide except in the European Union, the United Kingdom and South Korea; with an audience of over 100 million users per month, a separate license is required. VC-Attention is not a new model but a method for accelerating the attention layer in existing video generators. It does not change the architecture and does not require fine-tuning: V-Smooth reorders value tokens and quantizes the residual after subtracting the block mean, while ExpCast-FP8 replaces the FP32 exponential in softmax with a single fused operation. For owners of RTX 5090 and RTX PRO 6000, this is a potential way to lower memory requirements and speed up generation, but without published code the method cannot yet be tested on your own hardware.